code/superpoint_ops/batch_gen_superpoints.py

"""Batch generate pycut superpoint.npy for all S3DIS rooms."""
import os
import sys
import time
import numpy as np
from scipy.spatial import cKDTree

SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
LIBCP_DIR = os.path.join(SCRIPT_DIR, "_cut_pursuit", "build", "src")
sys.path.insert(0, LIBCP_DIR)
sys.path.insert(0, SCRIPT_DIR)

import libcp
from lib_geo import (
    _normalize_xyz_np,
    _normalize_normals_np,
    _local_geom_features_chunked_np,
    _build_adj_graph_np,
    _edge_weights_chunked_np,
    _relabel_contiguous_np,
)

S3DIS_ROOT = "/mnt/data/AODUOLI/_work_biptv3/pointcept_framework/data/s3dis_official"


def merge_small_components(xyz, labels, min_size=50):
    counts = np.bincount(labels)
    small_mask = counts[labels] < min_size
    if not small_mask.any():
        return labels
    labels = labels.copy()
    large_mask = ~small_mask
    if not large_mask.any():
        return labels
    tree = cKDTree(xyz[large_mask])
    _, nn_idx = tree.query(xyz[small_mask], k=1)
    large_indices = np.where(large_mask)[0]
    labels[small_mask] = labels[large_indices[nn_idx]]
    return _relabel_contiguous_np(labels)


def generate_superpoints_pycut(
    xyz, normals=None,
    k_feat=10, k_adj=10, chunk_size=8192,
    normal_scale=0.25, lam=0.03, sigma=0.5,
    min_comp_weight=20, weight_decay=0.7,
    merge_min_size=50,
):
    n = xyz.shape[0]
    xyz_norm = _normalize_xyz_np(xyz)
    geom_feat = _local_geom_features_chunked_np(xyz_norm, k_feat=k_feat, chunk_size=chunk_size)
    feat_parts = [geom_feat]
    if normals is not None:
        nn = _normalize_normals_np(normals)
        feat_parts.append(nn * normal_scale)
    Y = np.hstack(feat_parts).astype(np.float32)

    src, dst = _build_adj_graph_np(xyz_norm, k_adj=k_adj, mutual=False, undirected=True)
    ew = _edge_weights_chunked_np(Y.T, src, dst, lam=1.0, sigma=sigma)

    components, in_component = libcp.cutpursuit(
        Y,
        src.astype(np.uint32),
        dst.astype(np.uint32),
        ew.astype(np.float32),
        float(lam),
        int(min_comp_weight),
        0,
        float(weight_decay),
    )
    labels = _relabel_contiguous_np(np.asarray(in_component, dtype=np.int32))

    if merge_min_size > 0:
        labels = merge_small_components(xyz, labels, min_size=merge_min_size)

    return labels


if __name__ == "__main__":
    out_root = os.path.join(SCRIPT_DIR, "outputs", "superpoint_pycut_all")
    os.makedirs(out_root, exist_ok=True)

    areas = sorted([d for d in os.listdir(S3DIS_ROOT) if d.startswith("Area_")])
    total_rooms = 0
    for area in areas:
        area_dir = os.path.join(S3DIS_ROOT, area)
        rooms = sorted([r for r in os.listdir(area_dir)
                       if os.path.isdir(os.path.join(area_dir, r))])
        total_rooms += len(rooms)
    print(f"Total: {total_rooms} rooms across {len(areas)} areas")

    done = 0
    t_global = time.time()
    for area in areas:
        area_dir = os.path.join(S3DIS_ROOT, area)
        rooms = sorted([r for r in os.listdir(area_dir)
                       if os.path.isdir(os.path.join(area_dir, r))])

        for room in rooms:
            done += 1
            room_dir = os.path.join(area_dir, room)
            coord_path = os.path.join(room_dir, "coord.npy")
            normal_path = os.path.join(room_dir, "normal.npy")

            if not os.path.exists(coord_path):
                print(f"[{done}/{total_rooms}] SKIP {area}/{room}: no coord.npy")
                continue

            out_dir = os.path.join(out_root, area, room)
            os.makedirs(out_dir, exist_ok=True)
            out_path = os.path.join(out_dir, "superpoint.npy")

            if os.path.exists(out_path):
                print(f"[{done}/{total_rooms}] EXISTS {area}/{room}")
                continue

            coord = np.load(coord_path).astype(np.float32)
            normals = None
            if os.path.exists(normal_path):
                normals = np.load(normal_path).astype(np.float32)

            t0 = time.time()
            labels = generate_superpoints_pycut(
                coord, normals=normals,
                lam=0.03, sigma=0.5,
                k_feat=10, k_adj=10,
                merge_min_size=50,
            )
            dt = time.time() - t0
            n_sp = int(labels.max()) + 1

            np.save(out_path, labels)
            print(f"[{done}/{total_rooms}] {area}/{room}: {coord.shape[0]} pts -> {n_sp} sp  ({dt:.1f}s)")

    print(f"\nDone! Total time: {time.time() - t_global:.0f}s")